RS368: Grok Bot & Hermes

[00:00:00] Speaker A: Claude is my least favorite app, for sure, and it's been my least favorite model. It's been our team's least favorite model. Everyone's on Codex now. [00:00:09] Speaker B: Really moved over the superpowers of it, is that you have to give it superpowers to get real value. To make it as useful as it possibly can be, it needs to be able to do anything, which obviously gives it bad consequences, potentially. [00:00:25] Speaker A: You and I have learned so much by just fucking around. [00:00:29] Speaker B: Yeah, yeah. [00:00:29] Speaker A: That these companies aren't. Because people are like, well, every time I hit enter, it costs, you know, 70 cents. You multiply that times, you know, 5,000 people every day. Like, that's a lot. [00:00:45] Speaker B: You and I, we don't want a boss. We want to be our own boss. [00:00:48] Speaker A: Yeah. [00:00:49] Speaker B: Boss is generally quite good for getting you to make progress and do the right things. [00:00:54] Speaker A: Yeah. Hello. Welcome back to Rogue Startups, joined once again by Colin Gray. Colin, Today we're talking bots. Bots and agents. We finally are getting into the good stuff. I feel like we've been dancing around the topic du jour for a couple weeks, huh? [00:01:11] Speaker B: Yeah, I think there was a. There was a few things came up after our chat about it. You were mentioning Grokbot. We've both been talking a bit about what we've been doing with building more automated systems, I think, isn't it? So, yeah, it'd be cool to catch up on what you're doing, get some feedback maybe on the systems I've been building too. And also one little update I think I'd love to share around a renovation. Not of property. Well, partly of property, but of a digital property this time around, too. [00:01:39] Speaker A: I think the first question when you talk about agents is like, open versus closed. That's kind of one of the big decisions, I think. Crock Bot, obviously closed. I think at this point, you could call Codex an agent. I saw something on Twitter yesterday. They're supposed to be coming out with, like, you know, they acquired openclaw and Peter Steinberger. They're supposed to be coming out with a Codex agent soon, which will very, very much be like a Grok, but kind of thing. [00:02:03] Speaker B: Yeah. [00:02:04] Speaker A: Surely Anthropic will do the same. So I think from like, a. An initial, like, strategy perspective, like, hey, do you go on someone's turf and just accept the limitations and the issues there, or do you go, I know you've played around with openclaw quite a bit. We're building a Hermes agent for Kastos and chose it very intentionally because it is open. We control Everything, it runs on my hardware. You know, I was one of the suckers that bought a Mac Mini like February of this year in the open claw craze and are finally putting it to good use. But how do you think about, yeah, open or closed? [00:02:39] Speaker B: I mean, I'm usually an open type of guy. Like I would in the olden days. The argument was always like, what platform do you build your audience on, wasn't it? And you always were advising people like, don't build it on Facebook, don't build it on Twitter, build it on something you own. Like, don't build on rented land and the like. But I don't know, I think the question is. Or the answer is a bit different this time around, isn't it? Because it's a bit more about. It's a bit more about sustainability and about reliability and security as well, I suppose, isn't it? I mean, that was always, that was the big thing with openclaw when it first came out. Tell you be interested to hear how Hermes is with that as well these days since you're building that. But I think it's different, isn't it? I think this time around it's more about. I want to be able to do the work and not have to worry about maintenance and it going down all the time. So I'm veering more towards the closed, the prepared systems these days. But yeah, tell me, why did you go with Hermes then? Because that's open, obviously. So you're obviously going that route instead. [00:03:39] Speaker A: Good question. Very real concern. I think security is the biggest, biggest concern because otherwise the benefits are all there. You own the data, you control the models, you control the cost. You can run local models. Spoiler alert. Going to be speaking with Heat and Shaw actually this afternoon in real time, but it'll be the next week for this podcast about local models and open source. So, like, you know, whatever, subscribe and come back if you haven't subscribed already. To me, the shareability and configurability of Hermes is what makes it a better candidate for a company agent. Because it's not just me. Right. Our whole team uses it and so we need a. A surface where everyone can use it. So it's in Slack right now, but you can imagine we're going to build our own little like Trello kind of system to interface with the agent. Probably quite a bit harder to do that with Grokbot. When we started doing this, it was before you could share bots. [00:04:36] Speaker B: Yeah. [00:04:36] Speaker A: But even then I think everyone's kind of like Slightly like splintering the same bot and customizing for their own use. [00:04:44] Speaker B: I've not looked into this yet. So you. So you've been using Dropbot more than I have. You can't then install that in a publicly available space like Slack or something similar. Is that not the case yet? [00:04:54] Speaker A: No, that's what I'm saying. You can't. Yeah, from their perspective, it's. The whole appeal of it is like you just download it, sign up and it's. It's wheels included, it's the model, it's the chat interface, it's the harness, it's all the connectors all already there. With Hermes, it's like. And I saw the, you know these guys talking on Twitter yesterday where like they're going to even strip down more of it to where like it'll be less kind of batteries included. Hermes will in the future where like everything is something you have to intentionally add probably to the point of like security and performance. But yeah, that was the decision for us is like, it's not a personal assistant, it's a. You know, what we want it to be more like is a shared instance of CLAUDE code. I think that's the better mental model. It's like a shared instance of CLAUDE code that we orchestrate from some kind of community interface like Slack or a trello board or something like that. [00:05:49] Speaker B: You know, every have been talking about. They build their plus ones, I think they call it. And the whole concept of having other, like building AI teammates and having them in Slack, there's something that probably doesn't work about that when it's closed. Like if you're just talking to somebody else's AI teammate. So you're in engineering, somebody else is in marketing, you try and talk to their created bot, their simulated presence or whatever you want to call it, and it's just like back and forth with that. I don't know if that necessarily works, but when it's put open, suddenly these things have their own personality, their own presence, and the person that actually runs them can see it. They can see the interaction. They can potentially feedback and make it better or you know, know, say. I don't know if I necessarily agree with that and give actually some of their own feedback. Others can as well. Suddenly it's become so much more powerful at that point, doesn't it? [00:06:42] Speaker A: Yeah, I mean, interestingly, I saw they shut down their plus ones program every day. [00:06:48] Speaker B: Did they? I didn't notice that. [00:06:49] Speaker A: Which like, you know, for context, plus one is managed hosting for OpenClaw, basically from every and every Deadshipper is like a publication, like a AI publication. They do a bunch of training and stuff. I pay for it. It's one of the few things I pay for in the knowledge world. Yeah, they're very good. That's why we chose Hermes. But a very, like, a very specific thing that I was wrestling this week is you. You've used OpenClaw a lot. You know that it's, I'll say, brittle, right? It can break. You just send it one thing and it updates some kind of harness setting and the whole thing kind of breaks. Hermes is, like, slightly better, but. But, like, not a lot. So initially it was like, only I can talk to our Hermes agent. That's how it was initially in Slack. And I was like, that's dumb because, like, our team needs to be able to use it. So the gap between it's only me and it's literally everyone, and everyone can do everything. That's basically the step. It's either you or it's everyone, including settings and updating skills and updating the core and all this kind of stuff. So did a fair bit of work this week on, like, hey, how can we permission, essentially channels or profiles or users to be able to. Or not to be able to do things? But then it gets tough because, like, hey, the dev team needs access to the cli. How can we give them access to the CLI to spawn, like, cursor instances to do dev work, but not, do, you know, destructive stuff via the cli? [00:08:20] Speaker B: Did you get very far with it? Was it possible to do that? [00:08:23] Speaker A: Yeah, I did and actually, like, chat chatted with the Hermes team on Twitter about it. They gave me some, like, concrete things and some kind of soft things, but it's just like, this is the nature of, like, open source. And, you know, Hermes is a platform, you know, it's not a product. I would say, like, it's kind of like. It's just like WordPress, which. Which kind of fries my ass that I'm using it. But, like, it's like the starting point, right? It's getting you, like, eight steps down the road, and then you have to work on it the rest of the way. [00:08:53] Speaker B: It's also the classic sort of catch 22 in a way of Openclaw, when it first came out, which is the superpowers of it, is that you have to give it superpowers to get real value, to make it as useful as it possibly can be. It needs to be able to do anything, which obviously gives it bad consequences potentially. So as soon as you open that up to your team, then maybe that's even worse. But I don't know that really. That does excite me in a lot of ways to think of like being able to have assistants like that in my old team. Like if we could have had, you know, built up a marketing assistant that knew our content, it knew our campaigns, that knew our approach and you could just say to it like we just published an episode, go and do the standard repurposing plan and then it could set that all up and set up drafts and things for us to then go and review and edit and the team could do that and see it in progress as well. And then same for engineering and same for growth and sales and all sorts of areas that I love that idea of being able to work with a team. I think there's so many little things, hard parts that, that could really grease and get you moving. [00:09:58] Speaker A: Yeah, I mean I, I think the reality so far is like, yeah, like our agent, you know, created its first PR this week. It created its first bit, you know, bits of marketing and they were pretty average. You know, both, both of them were right because like our manual system has a lot more process and guardrails and context built in. And so we're, we're kind of like what I tell the team is like what we basically have to do is take all the stuff we're doing in cloud code and Codex right now and put it into Hermes and have it work like we work manually. It's just not the same yet. But totally like that that's the vision is like every, literally everything goes through this harness. [00:10:36] Speaker B: So that'll be running on your Mac Mini in your house. Your team will be accessing it from Slack or potentially you mentioned a kind [00:10:44] Speaker A: of Trello, like yeah, Slack today. Some kind of Trello esque kind of control panel in the future maybe. [00:10:51] Speaker B: Are you still using Grokbot then for personal stuff? Yeah. What's it turned out a few weeks later? What's it turned out as? The actual daily weekly drivers. [00:11:02] Speaker A: Yeah, I feel like a caveman when I use Grokbot. I mean I have it set up to like for Kastos like Chuck's help, scout a couple times a week and lets me know of anything that came in that like I should get involved in. Same with like Sentry. Same with like GitHub and like you know, dev stuff that's moving forward. So it's almost like a standup bot that's pulling information from our systems instead of the team reporting it it does the same with my inbox, both for like my personal and for my MyCastos inbox. The challenge is I don't use much of that to do stuff. You know, like, I get the ping and it's like, hey, here's your inbox. Here are the three things that need your attention. And I see that and then I go to Gmail. So like, I'm just not, I'm just not as effective as I should be. And exactly what I said of like, if everything routes through this harness, then it gets better. If I just. If it's just a notification system, then it's going to stay dumb. And that's my fault. [00:11:59] Speaker B: Yeah. So do you plan to change that then? Or is actually most of that going to just go through your new Hermes? I mean, if you're investing in this Hermes agent, making it actually super robust, secure, useful, knows all of your context. Would you shift it over there instead, even for your personal elements, or is that part of the segregation? [00:12:19] Speaker A: Yeah, both. Right. Like I, I definitely won't have the Hermes going through my personal stuff, but yeah, like the Help Scout and the Sentry and the GitHub. Yeah, you know, stuff sure enough should be going through the Hermes agent. I mean, maybe to this point, it's just so easy to do it in Grokbok. You just literally say, hey, connect to Sentry, give me a digest twice a week of what's going on there and then create issues for the three biggest looking problems that you're seeing errors on. Literally. That's literally it. How about you? Like, how are you? You're using both, right? Yeah. [00:12:56] Speaker B: Yes, in a way, I've actually, I've kind of come off most of it. I do have an open claw that still runs. I've basically relegated it to a note taker. It's a librarian, so I quite. I use it as a CRM actually. It's one of the most used parts. So I was down in Dundee this week, actually, met a few old contacts, great to catch up with them. And as soon as I come out of the meeting, I just talk into the Telegram and tell my openclaw to update my CRM and say this is a couple of things I learned about this person, what we should follow up with, maybe a next time to follow up, that kind of stuff. So the CRM side of things, actually I've found really useful. I'm not sure why exactly. I've found it more useful for that than other things, but that's just something I've stuck with, actually. And the other parts, I used to use it quite a lot as a task management. So to help me, because I'm working on like five or six different projects, I always had trouble like logging tasks and knowing where exactly to put them and logging information. And I would just let it do the filing, basically, which is what. What I mean by the librarian aspect. So instead of having to have seven or eight different repositories, it just puts it in the right place. I've kind of fallen off on that a little bit and just come back to using my to do app, so I have found it less useful. The thing I've ramped up though, is actually just using cloud code, honestly for just about everything, including updates, because I'm in there all day, every day anyway with my building. So, like whether it's the knowledge work stuff like working on marketing workflows, working on new content, or whether it's coding and building apps and things like that, I'm in there anyway. And so I've built these. I've got a daily brief that comes through every day, which looks at my calendar, looks at my inbox, looks at all the other kind of tasks. It helps me triage tasks, figure out what's most important that day, sends me that by email, comes into my inbox in the morning. And that's kind of taken over a lot of what I used to use the Open Claw for. So I have a feeling I'll probably even change the CRM over to that in the near future because Claude could do that too. [00:14:58] Speaker A: Yeah, something that I do along those lines is I use granola to record a lot of calls and it's automatically pulling into my Open Claw. Yeah, every call that it check checks a couple times a day, it pulls all the calls in and. And it updates its context of initiatives that I have based on what's going on. So like, for my coaching, record all the calls with everyone's permission and it pulls those in and then gives me a briefing based on. So like an email summary to the clients and then a prep message before my next one. So like, hey, last time you talked to Billy Bob, you talked about this and this and this. Make sure you touch on these three things that, that were on their plate. So, yeah, I think that's kind of a good example of like the thing I tell our team is like, the goal is like, remove cognitive load. Just like you're saying, like, you don't want to have to go into CRM. Remember all the shit that just happened in that meeting and blah, blah, Blah. You just want to like plop something into cloud code that is just easy so you don't have to think about it. And then all the downstream stuff, like works. I think that's like the mental model for me. [00:16:09] Speaker B: Yeah, yeah, it has to just work. That's the trouble. It has to not get in the way of the actual work, which is so often what the systems we build do. [00:16:17] Speaker A: Yeah, yeah. Well, what about like, if you're using cloud code as your driver for that? What about like mobile? Do you have it paired with your. And are you on a desktop because it's always on or like this thing? I. Because I don't sit at this desk maybe half the day and so I'm on the go all the time and do a lot of work from my phone or different places, so I can't really count on something on my computer to be the driver. [00:16:42] Speaker B: Yeah, so I've done a couple of things with that. One is I actually use Notion as the information logging spot. So I've got Notion linked up with CLAUDE via mcp. So when I'm using CLAUDE code, it creates like a build. It creates a build board for all my projects. So I can always see the state of the projects. I can see the to dos, the in progress, the complete. So I can always just go into Notion and see that. I know a lot of people do that in just markdown, for example, like you've just got markdown files, logs that way. But I like having it in Notion partly because it's more readable to me, it's just a bit more visual and I can go in and I can make edits really nicely. But the other part is that that then means that I can just open up a normal CLAUDE chat and that same CLAUDE account is linked up via MCP to the same Notion. So actually I can just drop in CRM updates straight into a normal Claude chat, like any old Claude chat, and it can go and check that it can update it and so it works really nicely that way. So the kind of knowledge work side of things is mostly taken care of in terms of syncing between normal CLAUDE and Claude code. The other part is the development and I did a lot of work recently actually to make sure that all of the projects that I care about are fully all GitHub synced. So I can. I've been using the Claude code cloud. Don't know what the actual term for it is. What is it? Do they call it someone in particular? [00:18:06] Speaker A: But there's a. Yeah, no, I know you're Talking about. Yeah, yeah. [00:18:08] Speaker B: Claude code, cloud mode. Claude code cloud mode. That's good. Tongue twister. [00:18:13] Speaker A: That's a lot. [00:18:14] Speaker B: So I've been using that quite a lot now so that I can actually, as long as I am confident and I'm usually confident in this. As long as I'm confident that I have committed and pushed everything from the latest sessions on my desktop, I can just open up a cloud session on my phone and actually just like do some updates that way and it syncs with the same GitHub and then next time I'm on desktop I just synchronize it again. So I found that actually really useful. So both for like app based sessions, but also knowledge work stuff. [00:18:44] Speaker A: And so do you have like a routine or a script or something that like when you fire up Claude it automatically, you know, pushes and pulls and syncs and all that with GitHub or do you have to remember to like when you start a session, oh, pull this down, merge all that kind of stuff? [00:18:57] Speaker B: Yeah, I have not yet actually. So yeah, that's something I have on the list is to make sure that the sessions. Yeah. When I open up the desktop version, it does go and pull it. Yeah. Because I haven't done that yet. [00:19:07] Speaker A: Yeah. You know, on, on this topic, like I just opened the Claude desktop app and I, I did a video yesterday about Grokbot and Cursor projects which we like. We can get into Cursor projects, but I haven't used Claude code in like three weeks because like the app, like for me just the app of Codex and, and Cursor are just so much better of a place to work. Like I don't like working in the terminal. I like, I mean it's this, it ends up being the same thing like Codecs or Cursor is just a, a chat interface but you're able to like view and preview and edit docs right in the app. I don't have to have like VS code opened up where it's like a little bit in the bottom, then the images are like, it's on the top. I think just in terms of like the way you interact with. I'll call them like agents that you need to be at the computer for. Right. Like Claude code, Codex, Cursor. Claude is my least favorite app for sure. And like it's, it's been my least favorite model. It's been our team's least favorite model. Everyone's on Codex now, really on our team and they're using it at least half the time depending on the person and like, what they're using. But yeah, super competitive space these days. [00:20:22] Speaker B: I keep hearing this and I keep thinking I need to actually switch over some stuff to Codex or anywhere else, to be honest. But I don't know, it's not been letting me down. I'm on the kind of middle max plan on Claude and I'm using Fable for planning. I'm using that to create like, you know, documents like project specs and things like that. And then I'm using Opus 5 to actually build the things and I'm finding that's a really good balance to actually, you know, create something without using up too many tokens and build it all. But I don't know, I hadn't even come across. Have you come across or tried Google's one yet? Anti Gravity? [00:20:58] Speaker A: Not since it first came out. [00:20:59] Speaker B: Yeah, no, it was funny because I was talking to somebody else earlier on that's doing a lot of building. They run an agency, actually, and they're converting their agency over to AI development as opposed to the old school version. And they've come to antigravity as their main sort of method for this. So I didn't know anything when he first said the name, I didn't even know what it was. So I looked it up and realized that it's Google's kind of agent model, I suppose, if you want to call it that, but they don't. But you can even use all the other models as well. Like they've got Opus in there and they've got a version of Codex in there as well, I believe so. It's a funny model, but. But he swears by that too. I think the main thing with that is it's actually super fast. Like, the main thing is it's just speed because it works with the. The latest Gemini model, I believe it is, and it's just like faster than all the rest. It's not too far behind on thought and reasoning and it's pretty close on coding, but it's just like 10 times as fast. So I think that's a big advantage. [00:22:00] Speaker A: Yeah, I mean, I mentioned, like talking to Heaton Shaw this afternoon. I think this is one of the bits of discipline that we all are going to need to get into soon. Is like, probably don't need to run Fable or Astra or even Opus for a lot of things we do. If we're not. And no, I don't mean this offensive to you. If we're not lazy, you know, if we're not trying to say like, hey, here's a bunch of stuff like go figure it out. But like, if we have skills and evals and systems and a real, you know, deterministic, like defined process for stuff, I don't think you should need the most advanced model. [00:22:39] Speaker B: Yeah, for sure. I went through I kind of painful process of this recently actually, because I've got my video workflow that I'm using for our golf channel just now. So any golfers out there, go and check out Both Sides of Par, the podcast. But I've been building a way to use all I grab too much video on the golf course. Craig. I think I must annoy my partner sometimes. Like, I'm just trying to take videos of just about every shot. I don't think I slow it down too much. But I've got all this video by the end of a round and I've been building loads of workflows to be able to turn that into decent stuff, whether it's, you know, taking shorts and adding typography or putting a shot tracer on a golf shot or putting together two or three shots into one log of an entire hole and putting a scoreboard up at the top. So I built little workflows to do all of those and long form as well and creating thumbnails for them too for these videos for YouTube. But all of these I've had the luxury over the last six months of having, I think I got so through a startup program I was part of when I was running Alitu. I got about £2,000 worth of anthropic credits, API credits. So that was sitting in my anthropic account. So I just went absolutely wild on this. Like, I was just like building all these processes that used everything and anything. And I didn't care about the models because I was never running out of tokens. I was. It was somebody else's money, so I didn't have to care. But recently I just discovered last week this whole workflow broke. I was like, what's going on? Something's going wrong. Looking at the errors and suddenly realized that my AP account, API account was empty. So turned out this money I'd been given had an expiry date on it, which I suppose makes sense. Can't keep that forever. But it meant that suddenly I had to put some money in the account and then continue to use this workflow that I had built. And so I ended up spending a day or two just reviewing how much every one of the little processes I had built cost, and therefore realizing that this was unsustainable. For the most part and downgrading them all. But that was exactly where I came to Craig. It was like I realized that nearly everything that I was doing could be downgraded at least one or two levels and still run perfectly fine. So there was even the. I thought the image generation was still going to cost me quite a lot, but I even downgraded that a level, I think in terms of the latest model and particularly the analysis I was doing analysis. So doing Shorts takes a script of 60 seconds and analyzes it for impact and does this for 20, 30 shorts all across a video to try and determine which five, let's say, are worth putting time into. I had that running on Opus, I believe, and I got no less good results by putting that right down to. I think I tried it on Sonnet. Still great. I think I even tried it on one of the Flash or something. I can't remember the name of it now, but it was a really low level model, cheap as chips and it was still pretty good. So yeah, you're spot on. So many things that we do that you don't need that kind of all purpose model for. [00:25:38] Speaker A: So I have a question. I've been wanting to do this for a long time, so I. The Mac, I got the cheapest Mac Mini you can get. It's 16 gigs of RAM, which you can't run anything. Any good model on that I've had in the cart for a couple of weeks. Mac Studio, which is the Mac Daddy, it's 128 gigs of RAM. [00:26:00] Speaker B: How much are they these days? Is it like five grand or something like that? More than that? [00:26:04] Speaker A: Yeah, it depends. The M4 version, which is like a one generation old Mac Studio, is about 6,000 for 128 gigs of RAM and like a two terabyte hard drive. [00:26:15] Speaker B: Okay. [00:26:16] Speaker A: Interestingly, as we're doing this, ChatGPT just took over my browser and is analyzing my YouTube channel. So pretty crazy. It's. It's like on a. It's on a schedule. I kind of want to do it because I think it would be pretty cool. Like a lot of what we want to do at Kastos could be run on a good local model. So you need about a hundred gigs of RAM to run Quinn 3.8 which is like the good local model these days. It's on par with like somewhere between Haiku and Sonnet probably. But if you think about like what do we use this for? It's like I want to write content, I want to do stuff with support, analyze like One of the big things we do right now, we're just in like, evaluation mode. On support is like twice a week, evaluate all the tickets that we've had come in that we've replied to, and evaluate those versus our knowledge base and see if there are areas in our knowledge base that we can improve based on questions that customers are asking. Not super complicated, you know, if you, if you try to like one shot, that more complicated. If you break that down into three or four steps, a simple model, you know, could take care of it. So anyways, like, I really just want to buy this because. Because I just do. The challenge I have is we're running our Hermes agent through my Codex plan and have not run out of credits yet. And that includes all of my own Codex use. So it's like, gosh, for a hundred bucks a month, the company basically runs its AI and I run mine. Like, that's 50, 60 months of at current, you know, capacity. [00:27:54] Speaker B: That's so much more generous than the anthropic plans, I think. I mean, I run out of my Max plan all the time. [00:28:00] Speaker A: Yeah. Oh, yeah. [00:28:01] Speaker B: Unless I am just completely wrong modeling it like we just mentioned. [00:28:05] Speaker A: No, I mean, I think. Yeah, I tell. I tell, like my, my consulting clients, like Claude, you know, call it two, three, five times more expensive per amount of work done than codex and like 10 times more than Cursor and Groq. You could never run out of your cursor plan, basically. I think. [00:28:27] Speaker B: Yeah, that's a good advert. [00:28:29] Speaker A: Yeah. Yeah, it's pretty awesome. [00:28:32] Speaker B: So, yeah, you had a question about that. But let me ask you something quickly first. You talking about a 6,000 pound or $Mac studio? There's quite a lot of AI specialized sort of mini computers coming out now, though, isn't there, for 2, 3000, maybe even less 1 or 2000. You're not tempted by that? Or is it really just the big shiny Mac Studio that you want? [00:28:53] Speaker A: I'm not aware of this. Like, what's an example of a specialized [00:28:57] Speaker B: computer like that They're PC builds, like, often in little cases. So they're kind of akin to a Mac Mini in many ways, but they're specially designed to run local models, so they'll just be chock full of RAM and a GPU or two, essentially. But you're seeing like, quite, supposedly quite powerful systems for. Yeah, a fair bit less than a Mac Studio and a lot smaller as well, like something you could carry around. This is one of the things I thought was quite interesting, like not a laptop, but a Mac mini style of thing that you could chuck in your bag to take with you to a CO working space or something similar. So yeah, there's a few of them around. It's an interesting sort of outcome of the local model or the AI development type approach. [00:29:41] Speaker A: Yeah, I mean I'm in the very early stages of working on a business idea around this right now because I think that whether it's local or just private AI and I'll differentiate the two of like local, it's like literally sitting here on my desk, private is like, it's a local model running in a CO located data center or something, but it's running an open weight model that's only accessible to, you know, whoever pays for it. I think there's a massive market for this and there will be more and more. I think one of the risks and I think there's like a hundred percent chance of it and we saw, we've seen it already with Anthropic is like, you know, on Anthropic, this is so fucking crazy. You and I are on personal accounts, right? And so we get like a seat and a bundle of credits. If you're an enterprise customer, you just get a seat and everything is API credits. Every time you chat, every time you run cloud code, it's API credits. [00:30:36] Speaker B: Which are a lot more expensive, aren't they? [00:30:38] Speaker A: I believe 2100 times more expensive. Yeah. [00:30:42] Speaker B: That much. Yeah. Because you often hear we are being heavily subsidized just now as our max plans. Yeah. [00:30:49] Speaker A: Yep. It has to be that either. And this is me kind of like convincing myself this is a good idea. But like fear around these platforms, you know, the frontier model, companies being compromised, what's really going on with your data and are they actually training on it when they say they're not, or just the cost? The cost and the sovereignty of your AI compute over time. I. There has to be a part of the market that will go this way. [00:31:15] Speaker B: Yeah, yeah, I think you're right. I looked at it and I wondered about it, but it's just so cheap and easy just now for the amount of. Yeah, it kind of, it was in my head, I was thinking, so if we're being so heavily subsidized just now, whereby we're paying £100amonth for a thousand pounds worth of credits, let's say that's like if I work on that for a year, then I'm getting, you know, it's 10k of funding for me for my business from Anthropic. I might as well use Some of that take something back from all the money they are making out of everyone. [00:31:47] Speaker A: But would you pay $1,000 a month for Claude coat? [00:31:51] Speaker B: No, not right now. No. Not for what I'm using it for. [00:31:54] Speaker A: Yeah. Oh, okay. If you were running Alitude still. [00:31:58] Speaker B: Possibly, possibly. But I mean at that point you really have to do the calculations, don't you? You have to work out like, is it actually paying back at this point? Like, is it cheaper than an extra developer? Is it more effective than an extra developer? Compare it to employees, compare it to, you know, the outcomes, the return on investment which then becomes a whole bully in itself which you don't really want to have to deal with. So then there's not only friction around, is it paying off? There's friction around. I know it's so expensive that I would need to treat it very seriously and I just don't have the headspace to do that. So I don't know. I think it would put a lot of people off. [00:32:34] Speaker A: Yeah, it's interesting. Like, you know, I think you and I are just like, oh, I'm going to do this thing, I'm going to create this website, I'm going to build this app which like, I know, like I have so many, if I look at GitHub, I have so many repos that have literally never seen the light of day. [00:32:47] Speaker B: Yes. [00:32:48] Speaker A: And that's fine if you're a company and you're spending API credits for everything that doesn't happen. And I think one thing that's interesting about that, because they're like, you know, cost conscious even if they have generous like allowances and initiatives to do more with AI. One thing that I think is interesting about that is you and I have learned so much by just fucking around. [00:33:06] Speaker B: Yeah, yeah. [00:33:07] Speaker A: That these companies aren't. Because people are like, well, every time I hit enter it costs, you know, 70 cents. You multiply that times, you know, 5,000 people every day. Like that's a lot. [00:33:20] Speaker B: There's almost. Yeah, there's almost the other effect as well, which is that I feel, and I'm sure you're the same. I've learned more in the last three months because I've been on this limited plan. What I mean by that is I'll find myself checking where I at, how much usage I have burned so far of my five hour window and my one week usage. And if I haven't got close to it yet, I'll deliberately just fire off a few weird and wonderful experiments just to use it up because I feel like If I'm not using my limit, if I'm not hitting my limit, I'm not getting my value for money. So there's all this. There's incentive to experiment, to do funny things that might not work, to just try stuff because it's there as well. Which I don't know whether that's a great thing or a bad thing, but I've certainly tried things that I definitely would not have under a usage plan. [00:34:13] Speaker A: Yeah. [00:34:13] Speaker B: One question that I had coming out of the bot, stuff that I had been looking at over the last few weeks is the whole model, especially with Grokbot, where people are moving towards a kind of chief of staff type plan, which then coordinates for you. It's something that quite attracts me because I'm getting a little bit. I'm building one app in particular at the moment, mainly for fun, but I've got a kind of eye on whether I could turn it into something in future. It's a kind of fantasy league type app, but for not sport, for other things. I've got some really interesting ideas around where it could be used. Mainly it's for use with me and my friends, but in building this, it's been great fun building out a whole bunch of stuff. It's one of the more complicated things that I've built with AI so far. And I've often had two, three, four threads open at once and then kind of wanted to break them into a few separate tasks as well. And therefore I end up. I'm not a developer. I've watched developers work. I know a work tree is versus a branch, all this kind of stuff, but still I get completely lost and I get a little bit burnt out coordinating it all as well and seeing where it all is in parallel. But the whole chief of staff model seems to be quite good for managing that type of thing. And then you bring in the fact that it helps you manage priorities across tasks as well. Potentially some schedule tasks in there too. Is that something you're using just now? [00:35:31] Speaker A: That's the idea with Grokbot, is everything runs through the chief of staff. I have two, one for Castos and one for personal stuff. Just because they each have separate, like, permissions, which I think is healthy. [00:35:43] Speaker B: Yeah, yeah. [00:35:44] Speaker A: I think in concept, what you're saying is true. The fear that I've seen is if you pile, and this is probably not unique to Grokkbot, but to Hermes or openclaw or, or whatever, if you're firing seven things at an agent at once, that things can get like, dropped. You Know if you're like, hey, go build this thing and check, help scout and write marketing content and do all this, from what I've seen that that's when it can just drop something. [00:36:11] Speaker B: Okay. Yeah. [00:36:12] Speaker A: So I don't know what the answer there is there. Again, maybe this is where like, is this like, open or closed? Like, they'll probably just figure it out. Whereas if it's Hermes, you're like, oh, gosh, like, we have to figure out how to get like a queuing system or something like that. [00:36:27] Speaker B: Yeah, yeah. [00:36:28] Speaker A: I'll say two things on the, like, actual development side, like, this video I did yesterday on like cursor projects. And I don't mean to like be repping cursor, but it is more like that, where it is a multi agent orchestration interface. It's a single chat that then all by itself spawns sub agents to do things. And so you can imagine for your app you would have a place to go to do everything. So it's like, hey, go do this. And it's going to go fire off an agent to do this. Hey, go build this thing. Hey, go do QA on this. Hey, go do marketing on this. Hey, build me a, you know, outreach list and interface with instantly or whatever. Like, whatever. Like, it's closed. But like, I think the, the difference there is like, the what I came to in my, in my kind of review of this is like, Crockbot is a personal assistant. It can be for team, it can be for work, but it's like to help you do more work, cursor projects is to like help you build stuff. [00:37:28] Speaker B: Yeah. Okay, that makes sense. [00:37:30] Speaker A: The one. And then this is for kind of everybody. The one thing that I haven't looked at is amp. I hear our mutual friend Brian Castle raving about it. And so, like, I trust him. And I would check out amp. It seems very similar in that it's cloud based, it's multi agent, and it's model agnostic. So you can use any model. [00:37:48] Speaker B: Yeah, I've heard a few people raving about it too. I've got it on my list to have a look at. I think the combination between that and Codex apparently is pretty powerful. [00:37:55] Speaker A: Oh, interesting. Okay. [00:37:56] Speaker B: Yeah. [00:37:57] Speaker A: So Codex can spawn amp kind of. [00:37:59] Speaker B: I believe so. Yeah. I've not looked into it, so don't hold me to that. But yeah, I've heard people talking about that for sure. One of the things when you're talking about chief of staff, one of the things that I think of, and this was something I hired a Real chief of staff for years ago in an actual business was stopping those projects that you mentioned a little while ago, never reaching light a day or stopping those projects even progressing. And also bringing back other things that have never quite finished. Like there's something around supervisor, manager, bot or agent of some sort that just constantly monitors the things you're working on day to day. And three days later, when you've got this thread that's still open in, you know, cursor and you're working away in Claude codes or whatever it says, by the way, just before you start in this new thing that you just asked me to do, you remember that thing you started like three days ago? That's kind of gone a bit stale by this point. You sure you don't want to just do that instead, rather than start this whole new thread? And it potentially could be a little bit annoying, of course, because it's going to try and stop you doing what you want to do. But equally that's exactly what you want a chief of staff to do in real life. Like if you actually run a company, if you want to achieve your goals, your aims, that's what it would be for. That's a big thing that I've been trying to build in the last little while, is a bot that helps me actually manage my week to week priorities, my day to day priorities. Tell me what to work on next. Because it's kind of as much as all of us like you and I, we don't want a boss, we want to be our own boss. Boss is generally quite good for getting you to make progress and do the right things. [00:39:39] Speaker A: As you're saying that like the, the two things that like instantly come to mind is like talking about permissioning, like it has to have visibility to literally everything, personal life, work life, across all of your kind of work surfaces, which they could be quite hard, especially if you're talking like cloud code, local development, you know, stuff and then you're talking to a cloud agent. Like the. Not, not that what you're saying is wrong, but like that, that's the, I think, limitation to it being successful. [00:40:03] Speaker B: Yeah, yeah. [00:40:04] Speaker A: And then the other one, and this is where I see the latest models from Anthropic and OpenAI. Being good in this respect for the first time is like I think we've talked about before, like it gets the big picture, it can go down a rabbit hole and then it can come back up and see the big picture again. I think both Fable and Astra can do this really well. You would have to Use to the point of like models. You would have to use a model like that to be like, ooh, let's get all of the context from all the stuff Colin's doing on a daily basis or whatever. Consider it in such a way that it can keep track of it over time because some of these things last forever, right. And then have the smarts to surface the thing that needs to be surfaced. I think that's a quite complex reasoning thing. [00:40:49] Speaker B: I think you're right, it absolutely is. [00:40:51] Speaker A: But it's just once a day, right? It's going to go, you're going to log all the stuff you're doing across everything and then once a day it's going to be like, cool, what happened? What do I need to let Colin know about? [00:40:58] Speaker B: Yeah, totally. And I think in some ways doing the plan ahead of time, like the old board model that we used to use, like you've got your board that runs once a quarter, once every two months, once every cycle you plan out, here's the plan for the two months, here's our goals and break it down into tasks. Then as long as that's the context, then actually some of the reasoning is not that tricky because does it fit into this? Does it move us towards this goal? No. Well, chuck it out. [00:41:24] Speaker A: Yeah, totally. [00:41:25] Speaker B: There's a few things I've been working on this past week which are completely talking of boards, it just popped into my head like proper traditional business. You were, I know you had a corporate business issue just before our call as well, which goes back to old school, how big old, big old corporate businesses run. But I had a couple of potential business plans around completely old school traditional type business approaches that be cool to come back to. Like, I know you have to talk about AI a fair bit in running business right now, but sometimes you're like, I don't know, can we just talk about like a building, you know, property or something like that? You know, something. We could come back to some updates around that in the next couple of weeks because I've got. There's a few interesting things. [00:42:06] Speaker A: Yeah, for sure, for sure. You teased like renovating a property like it sounds, sounds like a digital property. You gotta tell us about it. [00:42:13] Speaker B: Yeah. So actually just before I started the pod, well, as I started the podcast host, I had like four or five different things I was playing around with at the time. I was still working in a day job three times a week. So I had a couple of days a week to mess around with some digital businesses. And one of them I did with My dad actually. So my dad is a ex greenkeeper, used to look after golf courses, does a lot of lawn care. So he looks after people's gardens, but particularly grass. He knows grass like inside out. He's like world class grass guy. So we started a website called lawns4u. So lawnsforyou.com and I helped him build that first version of it. We put it was a WordPress site. We put a WooCommerce store on it and that was like 15 years ago. And he then took it on and just started putting content on it. So he's like written articles on it, kept it updated over the years, but it has barely been updated technically in 15 years. There's a few times I've gone on and like try to remember what the hell was going on with this site to fix a little bug that came up with a plugin that updated badly and took the whole site down as woocommerces want to do. But that is about it. But in the last few weeks I decided to see how an agent Claude code how Codex would be at helping me to overhaul this site and bring it back up today. And it's been amazing. It's been absolutely incredible what it can do when you just give it root access to the server, when you give it WP like a WordPress login and then access to like Google Cloud console, Google Analytics, basically all of the tools I used to use to run and grow a content site. So everything from like just literally getting rid of the bugs and making the thing look good right up to design side. So like redesigning the site, which hasn't been done yet. So if you go and look at that URL, don't think that's the new version. It's still in its terrible old state. But I'll be doing that soon to like huge SEO audits on the site. Like looking at 200 plus articles and saying, well here's like 12 different orphaned articles. Here's 14 that have broken links that used to link to this old product that's gone out a day. Here's 12 where the affiliate link here is out a day, here's 19. I think if we build this little extra article we could link them all to and build some real authority here because there's a huge gap that you've missed. And all these internal links and you know, all this stuff. Craig, like the SEO side of things that used to, I mean it's been absolutely crazy what it's been able to do, like with barely any input, not barely any Input. Sorry. I've been really directing it from a strategic point of view, but it just goes and does all of the actual legwork. So that's been really, really. And you're keeping it in WordPress debatable for now, yes. I thought that was a bigger change and that's definitely something I'd love to come back to, actually. What would you build a content site? Well, we could do that now if you want, but that's a big question, isn't it, what you build a content site in just now. [00:45:11] Speaker A: Yeah, I totally know where you're coming from. We. We did the, I'll say, like, research strategy and implementation part of that. Gosh, like at the beginning of the year probably for Castus, like, we. We built a cloud code project that would like, pull Google Analytics, search console and data for SEO, which is like an ahrefs kind of service, help us make decisions, create briefs, look at the serps, all this kind of stuff, and then write like quite good articles. But then it was like, okay, gosh, this is like in Markdown. We have to copy and paste it in WordPress. You could probably like MCP it in or something like that. I think WordPress is still like a good tool for like, a lot of people. You need some kind of cms. I think for most people our, like, the Kausta site is just markdown. It's just all markdown. It's Astro. [00:45:57] Speaker B: Yep. [00:45:57] Speaker A: And it's awesome. But that was a month of pretty intense time for me and we have a reasonably complex site. There's a lot going on, a lot of history and a lot of stuff to maintain. So, yeah, I think keeping it in WordPress for now is reasonable. [00:46:13] Speaker B: Yeah, I've heard a lot of good things about Astro. My son runs an Astro blog. So what was the pain that actually made you convert it then? What made it worth that month? [00:46:22] Speaker A: Exactly? Kind of what we're talking about. The ability to go from like monitoring, research, prep, writing, to publish all from Claude Code or Codex or whatever. [00:46:31] Speaker B: That's what I've been doing. So I've got Claude code actually with a login and it can go. It's updating articles with me. So I'm just doing it all straight in Claude code. I'm saying, right, what do we. What should we work on next? This Bowling Green Maintenance article hasn't been updated in 19 years, so let's update this one. And it pulls in the article just into literally the Claude code interface we both like. Ick says, here's some suggestions. I skim through. I think here's a few suggestions from me too. We work together a little bit on the edits and then it publishes the results. Yeah, directly. So did you find you couldn't do that easily? [00:47:10] Speaker A: Yeah, I guess the reality is I kind of just don't want to deal with WordPress anymore. [00:47:13] Speaker B: Well, that's fine, that's fair. [00:47:14] Speaker A: It's more of like religious discussion. Yes, yes. [00:47:21] Speaker B: No, that's fair. That's absolutely fair. And I could see, I could see me moving off WordPress for sure. I think the design part of it will be more of the issue for me in the end because I'm creating some really nice new designs for it right now. How I want it to look, how I want it to work. And I think a lot of this might not fit into WordPress. [00:47:37] Speaker A: Yeah, like for our, for ours we have like, you know, imagine, you know, the equivalent of like page templates and elements and stuff like that. So we have like a lot of tools, we're building a lot of tools. So we go in and just say like, hey, we want to build a tool. And maybe this is what would bring you over the edge, is like, like we have a tool that strips the audio out of a video file. It loads FFMPEG in the browser and does all of this in the browser and literally was one prompt in Codex. Hey, we want to do this, it needs to do this, blah, blah, blah, blah. Use the page template. We have link it from other places. It thought for about 10 minutes, gave me a preview URL on Cloudflare and I pushed publish. [00:48:15] Speaker B: Yeah, that is cool. But it'll start building tools for that audience. Yeah, absolutely. [00:48:19] Speaker A: Yeah. And like literally running like, especially if you're on Cloudflare, like it's Edge workers and stuff. Like you can do some pretty cool stuff without a database or a backend. Like there's literally no backend at all. [00:48:35] Speaker B: I'll put this on our docket for next time around, but I've been playing with a lot of different ways to the modern version of Lead Magnets. We did a bit of this with the podcast host actually was, you know, I mean it's just a kind of next stage of product based marketing, I suppose, but how easy it is to do now and the types, different types that you can do. Getting away from the old white paper or the PDF guide or whatever. But yeah, that's absolutely like the start of it, isn't it? Like how you engage an audience by building tiny little tools, but super useful, super targeted for them. And I've been doing one of them for our golf audience, actually, which has gone down really well so far. So we should come back to that. [00:49:12] Speaker A: Yeah. But I think that's probably the answer is like, when you know. Cause, like, man, I remember when I first started all this, I was like, how can you build a website without WordPress? Like, I just had no. I just had no concept. Yeah. [00:49:24] Speaker B: Yeah. [00:49:25] Speaker A: And. And I think when you go like, oh, I can do more than I can do in WordPress, that's probably when the decision is like, okay, WordPress is not the tool for me anymore. Cause I have things I want to do that it would just be way harder. [00:49:39] Speaker B: Yeah, it just holds it back. Yeah, definitely. [00:49:41] Speaker A: Yeah. I think that's like a good. Because it is a pain like that. It is. There is real pain to not using a cms. But you could. Like, one of the architecture decisions we made was like, hey, do we want a really lightweight cms? Like, for my personal site, I use. Oh, man, I can't remember. My personal site is a Next JS app with a very lightweight CMS built into it. Just for the blog. So the pages are pages. The blog is a lightweight cms. And so you could do that. [00:50:09] Speaker B: Yeah, definitely. There's a few nice lightweight ones around, I think, isn't there? I saw a couple of little blog systems, but. Yeah, that's cool. Let's come back to that. I'd like to know your thoughts of Framer as well, actually. I think there's a lot of talk goes around just now too. [00:50:21] Speaker A: Thanks, everyone for listening, and if you have recommendations for Colin and I for future episodes, give us a shout and we'll see you next time.